Custom rag output

[hummerichsander]

PR: Analysis Questions and Grammar handling

This PR adds analysis questions to the RAG job; analysis questions allow to specify additional questions the LLM should answer following a grammar the user can select. Furthermore, we introduce the Grammar model, that allows to specify additional grammars besides the usual yes/no grammar.

Changes

1. Added model AnalysisQuestion in the rag app. This model allows adding analysis questions to the RAG job.
2. Renamed model Question ← FilterQuestion in the rag app. This model holds the usual rag filter questions.
3. Adapt RagTaskProcessor class to handle both filter- and analysis-questions.
4. Added model Grammar in the chats app. This model captures the grammar string that allows to query a specific output. Besides the grammar string, one can provide a llm_instruction that is inserted into the message sent to the LLM. Example:

YesNoGrammar = Grammar(
    name="YES_NO",
    human_readable_name="Yes or No",
    grammar= dedent("""\
      root ::= Answer
      Answer ::= "Yes" | "No"
    """),
    llm_instruction="Answer each question in English faithfully with 'Yes' or 'No'."
)

5. The AsyncChatClient now requires a grammar argument in its ask_report_question and ask_question methods, which allows asking a yes-or-no question or a free-text question with the same method but with different grammar arguments.
6. Add CHAT_DEFAULT_GRAMMARS variable in the base settings. This list of default grammars is added to the database on compose-up via the create_default_grammars management command.